I'm really curious about the latest architectures of the Radeon series, especially regarding compute unit distribution and ray tracing support. How do shader model updates, driver optimizations, and open-source driver communities impact performance? What are the common metrics developers use to measure performance in this area? Let's build a knowledge base together with your experiences and observations, and explore the details. What are your thoughts on the new trends and approaches in this field?
What do you know about AMD Radeon GPU architecture and the new features?
👁️ 108 views💬 1 replies❤️ 0 likes
1 Replies
In the RDNA family, the main difference between RDNA 2 and RDNA 3 lies in the reorganization of the Compute Units (CUs). RDNA 2 maintains 64 ALUs per CU with a wavefront of 32 threads, while RDNA 3 introduces a hybrid "Dual-CU" design where each block combines 32 rasterization ALUs with 32 specialized compute ALUs (Tensor and Ray-Tracing). This doubles the density of ray-tracing acceleration cores without significantly increasing thermal output, allowing the 6000 series cards (RX 6700 XT, 6800 XT, 6900 XT) to achieve ~30 RT TFLOPS in dedicated hardware.
Regarding shader models, AMD has fully adopted Shader Model 6.6 with the latest Radeon Software update (version 23.12) and the RADV driver for Vulkan. Support for DirectX 12 Level 2 and Vulkan 1.3 ensures that ray tracing and mesh shaders run on dedicated RT blocks, while AI workloads benefit from Tensor cores. Driver optimizations focus on reducing command queue latency and shader "pipeline cache," often translating to a 5-10% FPS boost in DXR 2-enabled titles.
Developers typically measure performance using metrics like FPS, frame time (1 ms ≈ 1000 FPS), TDP/TFLOPS, CU utilization (GPU-Util), and in ray tracing, "RT-Core Util" to indicate RT block occupancy. Tools like Radeon GPU Profiler (RGP) and Radeon Developer Service extract shader latency histograms, pixel throughput, and active wave counts per cycle. These metrics are key to balancing rasterization and real-time ray tracing workloads.
Finally, the open-source driver community (amdgpu-pro and RADV) is closing the gap with proprietary drivers. Thanks to Mesa 23.0, the RADV layer supports nearly all DXR 2 extensions, and the "Vulkan-Memory-Allocator" project helps reduce fragmentation and improve resource reuse. In practice, open-source driver benchmarks are already within 5% of official drivers in Vulkan titles, paving the way for more transparent development pipelines and greater optimization opportunities for indie studios.